Casting Defect Detection — Augmented Multi-Class Dataset with 12 Defect Classes and Hand Overlay Augmentation Pipeline

This dataset provides synthetically augmented images of cast metal component defects across 12 classes for training and benchmarking computer vision models in industrial quality inspection tasks. WHAT MAKES THIS DATASET UNIQUE- 12 defect classes including the novel mixed_defects class (existing public datasets have 2-5 classes only)- Hand overlay augmentation using OpenCV GrabCut compositing — simulates real inspection scenarios, unavailable in any existing public casting dataset- Fully reproducible 3-stage Python pipeline included- 1,493 images | 145.77 MB DEFECT CLASSES (12 TOTAL)ok, cold_shut, crack, dent, flash, inclusion, mixed_defects, pinhole, porosity, scratch, shrinkage, surface_roughness PIPELINE (3 STAGES)Stage 1 — Defect Generation: 11 defect types synthesised onto real OK casting images using texture and geometry transforms via casting_defect_augmentation.py Stage 2 — Hand Overlay Augmentation: ALL casting images composited with ALL hand images using GrabCut background removal. 44 variants per combination (24 predefined + 20 random) via real_hand_overlay.py Stage 3 — Visualisations: Sample grids, intensity comparisons, and augmentation demos via demo_visualization.py FOLDER STRUCTUREdefect_dataset/├── ok/├── cold_shut/├── crack/├── dent/├── flash/├── inclusion/├── mixed_defects/├── pinhole/├── porosity/├── scratch/├── shrinkage/└── surface_roughness/ REPRODUCE THIS DATASETpip install -r requirements.txtpython run_complete_pipeline.py USE CASES- Binary defect classifier (OK vs Defective)- 12-way multi-class defect type classification- Anomaly detection benchmarking- Industrial data augmentation research- Transfer learning for manufacturing inspection QUICK START (3 lines of code)from tensorflow.keras.utils import image_dataset_from_directoryds = image_dataset_from_directory("defect_dataset", image_size=(224,224))